MiniMax just dropped a heavyweight AI agent — and then quietly rewired the rules around how you can use it.
What’s out: M2.7, a 230B-parameter Mixture-of-Experts (MoE) model that activates only ~10B parameters per inference, is now on Hugging Face and punching at the level of the best closed models. Benchmarks: 56.22% on SWE‑Pro (software-engineering tasks), ~57.0% on Terminal Bench 2, and an ELO of 1495 on GDPval‑AA (real-world knowledge work). That ELO is the highest among open-weight models and sits just behind heavy hitters like Opus 4.6, Sonnet 4.6 and GPT‑5.4. MiniMax also claims an unusual development story: an internal version ran 100+ autonomous self-optimization rounds, rewrote its own scaffold and emerged ~30% better — all without a human-in-the-loop.
What changed: shortly after the weights went live, MiniMax updated the license terms. Non-commercial uses remain free and unrestricted — research, personal projects and fine-tuning for private setups are unchanged. But commercial use now requires written authorization from MiniMax. In practice that means hosted services or commercial products deploying M2.7 must get explicit permission.
Why people complained: the community reaction on Hacker News and Hugging Face was immediate. MiniMax labeled the new terms “MIT‑style,” which many developers found misleading because the MIT license explicitly permits commercial use. Describing the change as “Modified‑MIT” while introducing a commercial-use gate created confusion and frustration.
MiniMax’s response: Ryan Lee, Head of Developer Relations, posted a detailed explanation. He said the move was driven by bad-faith hosting providers that had been deploying degraded or misconfigured versions of prior MiniMax models — wrong templates, aggressive quantization, or not even the real model — and leaving users with a poor impression that hurt MiniMax’s reputation. A fully permissive license, Lee argued, left them no leverage to push back. He invited feedback on edge cases and said the team would rather amend the license text than defend it, and that commercial authorization will be “fast and reasonable.”
Bigger picture: this is a break from MiniMax’s recent openness. M2 and M2.5 were released under MIT in Oct 2025 and Feb 2026 respectively. M2.7’s shift comes months after MiniMax’s Jan 2026 Hong Kong IPO (roughly $620M raised, with backers including Alibaba and Abu Dhabi’s sovereign fund). It also lands amid a broader trend: several Chinese AI labs that once emphasized open weights are experimenting with more restrictive licensing — Alibaba’s Qwen reportedly moved toward proprietary development, and Xiaomi released MiMo v2 under a closed-source license. The old shorthand — Chinese labs open, U.S. labs closed — no longer fits.
What this means for builders and crypto projects: M2.7 offers frontier-level inference cost savings (MoE, 10B active params) and strong benchmark performance, making it attractive for startups, on-chain/off-chain agents, and decentralized inference marketplaces. But teams planning to run hosted services or embed the model in commercial products need to secure authorization from MiniMax first. For research, personal use, and private fine-tuning, nothing changes.
Bottom line: M2.7 raises the bar technically and complicates the licensing landscape politically and commercially. Watch how MiniMax clarifies the license language and how hosting providers and the developer community respond — that will determine whether M2.7 becomes widely adopted in production or primarily used in labs and private deployments.
Read more AI-generated news on: undefined/news
